RaspberryPi3(Raspbian Stretch) or Ubuntu16.04/UbuntuMate + Neural Compute Stick(NCS/NCS2) + RealSense D435(or USB Camera or PiCamera) + MobileNet-SSD(MobileNetSSD)
【Notice】December 19, 2018 OpenVINO has supported RaspberryPi + NCS2 !!
https://software.intel.com/en-us/articles/OpenVINO-RelNotes#inpage-nav-2-2
【Dec 31, 2018】 USB Camera + MultiStick + MultiProcess mode correspondence with NCS2 is completed.
【Jan 04, 2019】 Tune performance four times. MultiStickSSDwithRealSense_OpenVINO_NCS2.py. Core i7 -> NCS2 x1, 48 FPS
【Nov 12, 2019】 Compatible with OpenVINO 2019 R3 + RaspberryPi3/4 + Raspbian Buster.
Measure the distance to the object with RealSense D435 while performing object detection by MobileNet-SSD(MobileNetSSD) with RaspberryPi 3 boosted with Intel Movidius Neural Compute Stick.
"USB Camera mode / PiCamera mode" can not measure the distance, but it operates at high speed.
And, This is support for MultiGraph and FaceDetection, MultiProcessing, Background transparentation.
And, This is support for simple clustering function. (To prevent thermal runaway)
【Japanese Article1】
RaspberryPi3 (Raspbian Stretch) + Intel Movidius Neural Compute Stick(NCS) + RealSenseD435 + MobileNet-SSD(MobileNetSSD) で高速に物体検出しつつ悟空やモニタまでの距離を測る
【Japanese / English Article2】
Intel also praised me again ヽ(ﾟ∀ﾟ)ﾉ Yeah MobileNet-SSD(MobileNetSSD) object detection and RealSense distance measurement (640x480) with RaspberryPi3 At least 25FPS playback frame rate + 12FPS prediction rate
【Japanese / English Article3】
Detection rate approx. 30FPS RaspberryPi3 Model B(plus none) is slightly later than TX2, acquires object detection rate of MobilenetSSD and corresponds to MultiModel VOC+WIDER FACE
【Japanese Article4】
RaspberryPi3で複数のMovidius Neural Compute Stick をシームレスにクラスタ切り替えして高速推論性能を維持しつつ熱暴走(内部温度70℃前後)を回避する
【Japanese Article5】
Caffeで超軽量な "Semantic Segmentation" のモデルを生成する Sparse-Quantized CNN 512x1024_10MB_軽量モデル_その１
【Japanese / English Article6】
Boost RaspberryPi3 with Neural Compute Stick 2 (1 x NCS2) and feel the explosion performance of MobileNet-SSD (If it is Core i7, 21 FPS)
【Japanese / English Article7】
[24 FPS] Boost RaspberryPi3 with four Neural Compute Stick 2 (NCS2) MobileNet-SSD / YoloV3 [48 FPS for Core i7]
【Japanese / English Article8】
[24 FPS, 48 FPS] RaspberryPi3 + Neural Compute Stick 2, The day when the true power of one NCS2 was drawn out and "Goku" became a true "super saiya-jin"
1. Summary
1.1 Verification environment NCSDK (1)
1.2 Result of detection rate NCSDK (1)
1.3 Verification environment NCSDK (2)
1.4 Result of detection rate NCSDK (2)
2. Performance comparison as a mobile application (Based on sensory comparison)
3. Change history
4. Motion image
4-1. NCSDK ver
4-1-1. RealSense Mode about 6.5 FPS （Synchronous screen drawing）
4-1-2. RealSense Mode about 25.0 FPS （Asynchronous screen drawing）
4-1-3. USB Camera Mode MultiStick x4 Boosted 16.0 FPS+ （Asynchronous screen drawing）
4-1-4. RealSense Mode SingleStick about 5.0 FPS（Transparent background / Asynchronous screen drawing
4-1-5. USB Camera Mode MultiStick x3 Boosted （Asynchronous screen drawing / MultiGraph
4-1-6. Simple clustering function (MultiStick / MultiCluster / Cluster switch cycle / Cluster switch temperature)
4-2. OpenVINO ver
4-2-1. USB Camera Mode NCS2 x 1 Stick + RaspberryPi3（Synchronous screen drawing）
4-2-2. USB Camera Mode NCS2 x 1 Stick + Core i7（Synchronous screen drawing）
4-2-3. USB Camera Mode NCS2 x 1 Stick + Core i7（Asynchronous screen drawing）
4-2-4. USB Camera Mode NCS2 x 1 Stick + RaspberryPi3（Asynchronous screen drawing）
4-2-5. USB Camera Mode NCS2 x 1 Stick + LattePanda Alpha（Asynchronous screen drawing）48 FPS
4-2-6. PiCamera Mode NCS2 x 1 Stick + RaspberryPi3（Asynchronous screen drawing）
4-2-7. USB Camera Mode NCS2 x 1 Stick + RaspberryPi4（Asynchronous screen drawing）40 FPS
5. Motion diagram of MultiStick
6. Environment
7. Firmware update with Windows 10 PC
8. Work with RaspberryPi3 (or PC + Ubuntu16.04 / RaspberryPi + Ubuntu Mate)
8-1. NCSDK ver (Not compatible with NCS2)
8-2. OpenVINO ver (Corresponds to NCS2)
9. Execute the program
10. 【Reference】 MobileNetv2 Model (Caffe) Great Thanks!!
11. Conversion method from Caffe model to NCS model (NCSDK)
12. Conversion method from Caffe model to NCS model (OpenVINO)
13. Construction of learning environment and simple test for model (Ubuntu16.04 x86_64 PC + GPU NVIDIA Geforce)
14. Reference articles, thanks
Performance measurement result each number of sticks. (It is Detection rate. It is not a Playback rate.)
The best performance can be obtained with QVGA + 5 Sticks.
However, It is important to use a good quality USB camera.
| No. | Item | Contents |
|---|---|---|
| 1 | Video device | USB Camera (No RealSense D435) ELP-USB8MP02G-L75 $70 |
| 2 | Auxiliary equipment | (Required) self-powered USB2.0 HUB |
| 3 | Input resolution | 640x480 |
| 4 | Output resolution | 640x480 |
| 5 | Execution parameters | $ python3 MultiStickSSDwithRealSense.py -mod 1 -wd 640 -ht 480 |
| No. | Stick count | FPS | Youtube Movie | Note |
|---|---|---|---|---|
| 1 | 1 Stick | 6 FPS | https://youtu.be/lNbhutT8hkA | base line |
| 2 | 2 Sticks | 12 FPS | https://youtu.be/zuJOhKWoLwc | 6 FPS increase |
| 3 | 3 Sticks | 16.5 FPS | https://youtu.be/8UDFIJ1Z4v8 | 4.5 FPS increase |
| 4 | 4 Sticks | 16.5 FPS | https://youtu.be/_2xIZ-IZwZc | No improvement |
| No. | Item | Contents |
|---|---|---|
| 1 | Video device | USB Camera (No RealSense D435) PlayStationEye $5 |
| 2 | Auxiliary equipment | (Required) self-powered USB2.0 HUB |
| 3 | Input resolution | 320x240 |
| 4 | Output resolution | 320x240 |
| 5 | Execution parameters | $ python3 MultiStickSSDwithRealSense.py -mod 1 -wd 320 -ht 240 |
| No. | Stick count | FPS | Youtube Movie | Note |
|---|---|---|---|---|
| 1 | 4 Sticks | 25 FPS | https://youtu.be/v-Cei1TW88c | |
| 2 | 5 Sticks | ⭐ 30 FPS | https://youtu.be/CL6PTNgWibI | best performance |
◯=HIGH, △=MEDIUM, ×=LOW
| No. | Model | Speed | Accuracy | Adaptive distance |
|---|---|---|---|---|
| 1 | SSD | × | ◯ | ALL |
| 2 | MobileNet-SSD | △ | △ | Short distance |
| 3 | YoloV3 | × | ◯ | ALL |
| 4 | tiny-YoloV3 | ◯ | × | Long distance |
Change history
[July 17, 2018] Corresponds to OpenCV 3.4.2
[July 21, 2018] Support for multiprocessing [MultiStickSSDwithRealSense.py]
[July 23, 2018] Support for USB Camera Mode [MultiStickSSDwithRealSense.py]
[July 29, 2018] Added steps to build learning environment
[Aug 3, 2018] Background Multi-transparent mode implementation [MultiStickSSDwithRealSense.py]
[Aug 11, 2018] CUDA9.0 + cuDNN7.2 compatible with environment construction procedure
[Aug 14, 2018] Reference of MobileNetv2 Model added to README and added Facedetection Model
[Aug 15, 2018] Bug Fixed. `MultiStickSSDwithRealSense.py` depth_scale be undefined. Pull Requests merged. Thank you Drunkar!!
[Aug 19, 2018] 【Experimental】 Update Facedetection model [DeepFace] (graph.facedetectXX)
[Aug 22, 2018] Separate environment construction procedure of "Raspbian Stretch" and "Ubuntu16.04"
[Aug 22, 2018] 【Experimental】 FaceDetection model replaced [resnet] (graph.facedetection)
[Aug 23, 2018] Added steps to build NCSDKv2
[Aug 25, 2018] Added "Detection FPS View" [MultiStickSSDwithRealSense.py]
[Sep 01, 2018] FaceDetection model replaced [Mobilenet] (graph.fullfacedetection / graph.shortfacedetection)
[Sep 01, 2018] Added support for MultiGraph and FaceDetection mode [MultiStickSSDwithRealSense.py]
[Sep 04, 2018] Performance measurement result with 5 sticks is posted
[Sep 08, 2018] To prevent thermal runaway, simple clustering function of stick was implemented.
[Sep 16, 2018] 【Experimental】 Added Semantic Segmentation model [Tensorflow-UNet] (semanticsegmentation_frozen_person.pb)
[Sep 20, 2018] 【Experimental】 Updated Semantic Segmentation model [Tensorflow-UNet]
[Oct 07, 2018] 【Experimental】 Added Semantic Segmentation model [caffe-jacinto] (cityscapes5_jsegnet21v2_iter_60000.caffemodel)
[Oct 10, 2018] Corresponds to NCSDK 2.08.01
[Oct 12, 2018] 【Experimental】 Added Semantic Segmentation model [Tensorflow-ENet] (semanticsegmentation_enet.pb) https://github.com/PINTO0309/TensorFlow-ENet.git
[Dec 22, 2018] Only "USB Camera + single thread mode" correspondence with NCS 2 is completed
[Dec 31, 2018] "USB Camera + MultiStick + MultiProcess mode" correspondence with NCS2 is completed
[Jan 04, 2019] Tune performance four times. MultiStickSSDwithRealSense_OpenVINO_NCS2.py
[Feb 01, 2019] Pull request merged. Fix Typo. Thanks, nguyen-alexa!!
[Feb 09, 2019] Corresponds to PiCamera.
[Feb 10, 2019] Added support for SingleStickSSDwithRealSense_OpenVINO_NCS2.py
[Feb 10, 2019] Firmware v5.9.13 -> v5.10.6, RealSenseSDK v2.13.0 -> v2.16.5
[May 01, 2019] Corresponds to OpenVINO 2019 R1.0.1
[Nov 12, 2019] Corresponds to OpenVINO 2019 R3.0
RealSense Mode about 6.5 FPS （Detection + Synchronous screen drawing / SingleStickSSDwithRealSense.py）
【YouTube Movie】 https://youtu.be/77cV9fyqJ1w
However, the prediction rate is fairly low.(about 6.5 FPS)
【YouTube Movie】 https://youtu.be/tAf1u9DKkh4
USB Camera Mode MultiStick x4 Boosted 16.0 FPS+ （Asynchronous screen drawing / MultiStickSSDwithRealSense.py）
【YouTube Movie】 https://youtu.be/GedDpAc0JyQ
RealSense Mode SingleStick about 5.0 FPS（Transparent background in real time / Asynchronous screen drawing / MultiStickSSDwithRealSense.py）
【YouTube Movie】 https://youtu.be/ApyX-mN_dYA
USB Camera Mode MultiStick x3 Boosted （Asynchronous screen drawing / MultiGraph(SSD+FaceDetection) / FaceDetection / MultiStickSSDwithRealSense.py）
【YouTube Movie】 https://youtu.be/fQZpuD8mWok
Simple clustering function (MultiStick / MultiCluster / Cluster switch cycle / Cluster switch temperature)
[Execution log]
USB Camera Mode NCS2 SingleStick + RaspberryPi3（Synchronous screen drawing / SingleStickSSDwithUSBCamera_OpenVINO_NCS2.py）
【YouTube Movie】 https://youtu.be/GJNkX-ZBuC8
USB Camera Mode NCS2 SingleStick + Core i7（Synchronous screen drawing / SingleStickSSDwithUSBCamera_OpenVINO_NCS2.py）
【YouTube Movie】 https://youtu.be/1ogge90EuqI
USB Camera Mode NCS2 x 1 Stick + Core i7（Asynchronous screen drawing / MultiStickSSDwithRealSense_OpenVINO_NCS2.py）
【YouTube Movie】 https://youtu.be/Nx_rVDgT8uY
$ python3 MultiStickSSDwithRealSense_OpenVINO_NCS2.py -mod 1 -numncs 1
USB Camera Mode NCS2 x 1 Stick + RaspberryPi3（Asynchronous screen drawing / MultiStickSSDwithRealSense_OpenVINO_NCS2.py）
【YouTube Movie】 https://youtu.be/Xj2rw_5GwlI
$ python3 MultiStickSSDwithRealSense_OpenVINO_NCS2.py -mod 1 -numncs 1
USB Camera Mode NCS2 x 1 Stick + LattePanda Alpha（Asynchronous screen drawing / MultiStickSSDwithRealSense_OpenVINO_NCS2.py）[48 FPS]
https://twitter.com/PINTO03091/status/1081575747314057219
PiCamera Mode NCS2 x 1 Stick + RaspberryPi3（Asynchronous screen drawing / MultiStickSSDwithPiCamera_OpenVINO_NCS2.py）
$ python3 MultiStickSSDwithPiCamera_OpenVINO_NCS2.py
USB Camera Mode NCS2 x 1 Stick + RaspberryPi4（Asynchronous screen drawing / MultiStickSSDwithUSBCamera_OpenVINO_NCS2.py）
$ python3 MultiStickSSDwithUSBCamera_OpenVINO_NCS2.py
1．RaspberryPi3 + Raspbian Stretch (USB2.0 Port) or RaspberryPi3 + Ubuntu Mate or PC + Ubuntu16.04
2．Intel RealSense D435 (Firmware Ver 5.10.6) or USB Camera or PiCamera Official stable version firmware
3．Intel Neural Compute Stick v1/v2 x１piece or more
4-1．OpenCV 3.4.2 (NCSDK)
4-2．OpenCV 4.1.1-openvino (OpenVINO)
5．VFPV3 or TBB (Intel Threading Building Blocks)
6．Numpy
7．Python3.5
8．NCSDK v2.08.01 (It does not work with NCSDK v1. v1 version is here)
9. OpenVINO 2019 R2.0.1
10．RealSenseSDK v2.16.5 (The latest version is unstable) Official stable version SDK
11．HDMI Display
1．ZIP 2 types (1) Firmware update tool for Windows 10 (2) The latest firmware bin file Download and decompress
2．Copy Signed_Image_UVC_5_10_6_0.bin to the same folder as intel-realsense-dfu.exe
3．Connect RealSense D435 to USB port
4．Wait for completion of installation of device driver
5．Execute intel-realsense-dfu.exe
6．「1」 Type and press Enter and follow the instructions on the screen to update
7．Firmware version check 「2」
Use of Virtualbox is not strongly recommended
[Note] Japanese Article
https://qiita.com/akitooo/items/6aee8c68cefd46d2a5dc
https://qiita.com/kikuchi_kentaro/items/280ac68ad24759b4c091
[Post of Official Forum]
https://ncsforum.movidius.com/discussion/950/problems-with-python-multiprocessing-using-sdk-2-0-0-4
https://ncsforum.movidius.com/discussion/comment/3921
https://ncsforum.movidius.com/discussion/comment/4316/#Comment_4316
1.Execute the following
$ sudo apt update;sudo apt upgrade
$ sudo reboot2.Extend the SWAP area (RaspberryPi+Raspbian Stretch / RaspberryPi+Ubuntu Mate Only)
$ sudo nano /etc/dphys-swapfile
CONF_SWAPSIZE=2048
$ sudo /etc/init.d/dphys-swapfile restart;swapon -s3.Install NSCDK
$ sudo apt install python-pip python3-pip
$ sudo pip3 install --upgrade pip
$ sudo pip2 install --upgrade pip
$ cd ~/ncsdk
$ make uninstall
$ cd ~;rm -r -f ncsdk
#=====================================================================================================
# [Oct 10, 2018] NCSDK 2.08.01 , Tensorflow 1.9.0
$ git clone -b ncsdk2 http://github.com/Movidius/ncsdk
#=====================================================================================================
$ cd ncsdk
$ nano ncsdk.conf
#MAKE_NJOBS=1
↓
MAKE_NJOBS=1
$ sudo apt install cython
$ sudo -H pip3 install cython
$ sudo -H pip3 install numpy
$ sudo -H pip3 install pillow
$ make install
$ cd ~
$ wget https://github.com/google/protobuf/releases/download/v3.5.1/protobuf-all-3.5.1.tar.gz
$ tar -zxvf protobuf-all-3.5.1.tar.gz
$ cd protobuf-3.5.1
$ ./configure
$ sudo make -j1
$ sudo make install
$ cd python
$ export LD_LIBRARY_PATH=../src/.libs
$ python3 setup.py build --cpp_implementation
$ python3 setup.py test --cpp_implementation
$ sudo python3 setup.py install --cpp_implementation
$ sudo ldconfig
$ protoc --version
# Before executing "make examples", insert Neural Compute Stick into the USB port of the device.
$ cd ~/ncsdk
$ make examples -j1【Reference】https://github.com/movidius/ncsdk
4.Update udev rule
$ sudo apt install -y git libssl-dev libusb-1.0-0-dev pkg-config libgtk-3-dev
$ sudo apt install -y libglfw3-dev libgl1-mesa-dev libglu1-mesa-dev
$ cd /etc/udev/rules.d/
$ sudo wget https://raw.githubusercontent.com/IntelRealSense/librealsense/master/config/99-realsense-libusb.rules
$ sudo udevadm control --reload-rules && udevadm trigger5.Upgrade to "cmake 3.11.4"
$ cd ~
$ wget https://cmake.org/files/v3.11/cmake-3.11.4.tar.gz
$ tar -zxvf cmake-3.11.4.tar.gz;rm cmake-3.11.4.tar.gz
$ cd cmake-3.11.4
$ ./configure --prefix=/home/pi/cmake-3.11.4
$ make -j1
$ sudo make install
$ export PATH=/home/pi/cmake-3.11.4/bin:$PATH
$ source ~/.bashrc
$ cmake --version
cmake version 3.11.46.Register LD_LIBRARY_PATH
$ nano ~/.bashrc
export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH
$ source ~/.bashrc7.Install TBB (Intel Threading Building Blocks)
$ cd ~
$ wget https://github.com/PINTO0309/TBBonARMv7/raw/master/libtbb-dev_2018U2_armhf.deb
$ sudo dpkg -i ~/libtbb-dev_2018U2_armhf.deb
$ sudo ldconfig8.Uninstall old OpenCV (RaspberryPi Only)
[Very Important] The highest performance can not be obtained unless VFPV3 is enabled.
$ cd ~/opencv-3.x.x/build
$ sudo make uninstall
$ cd ~
$ rm -r -f opencv-3.x.x
$ rm -r -f opencv_contrib-3.x.x9.Build install "OpenCV 3.4.2" or Install by deb package.
[Very Important] The highest performance can not be obtained unless VFPV3 is enabled.
9.1 Build Install (RaspberryPi Only)
$ sudo apt update && sudo apt upgrade
$ sudo apt install build-essential cmake pkg-config libjpeg-dev libtiff5-dev \
libjasper-dev libavcodec-dev libavformat-dev libswscale-dev \
libv4l-dev libxvidcore-dev libx264-dev libgtk2.0-dev libgtk-3-dev \
libcanberra-gtk* libatlas-base-dev gfortran python2.7-dev python3-dev
$ cd ~
$ wget -O opencv.zip https://github.com/Itseez/opencv/archive/3.4.2.zip
$ unzip opencv.zip;rm opencv.zip
$ wget -O opencv_contrib.zip https://github.com/Itseez/opencv_contrib/archive/3.4.2.zip
$ unzip opencv_contrib.zip;rm opencv_contrib.zip
$ cd ~/opencv-3.4.2/;mkdir build;cd build
$ cmake -D CMAKE_CXX_FLAGS="-DTBB_USE_GCC_BUILTINS=1 -D__TBB_64BIT_ATOMICS=0" \
-D CMAKE_BUILD_TYPE=RELEASE \
-D CMAKE_INSTALL_PREFIX=/usr/local \
-D INSTALL_PYTHON_EXAMPLES=OFF \
-D OPENCV_EXTRA_MODULES_PATH=~/opencv_contrib-3.4.2/modules \
-D BUILD_EXAMPLES=OFF \
-D PYTHON_DEFAULT_EXECUTABLE=$(which python3) \
-D INSTALL_PYTHON_EXAMPLES=OFF \
-D BUILD_opencv_python2=ON \
-D BUILD_opencv_python3=ON \
-D WITH_OPENCL=OFF \
-D WITH_OPENGL=ON \
-D WITH_TBB=ON \
-D BUILD_TBB=OFF \
-D WITH_CUDA=OFF \
-D ENABLE_NEON:BOOL=ON \
-D ENABLE_VFPV3=ON \
-D WITH_QT=OFF \
-D BUILD_TESTS=OFF ..
$ make -j1
$ sudo make install
$ sudo ldconfig9.2 Install by deb package (RaspberryPi Only) [I already activated VFPV3 and built it]
$ cd ~
$ sudo apt autoremove libopencv3
$ wget https://github.com/PINTO0309/OpenCVonARMv7/raw/master/libopencv3_3.4.2-20180709.1_armhf.deb
$ sudo apt install -y ./libopencv3_3.4.2-20180709.1_armhf.deb
$ sudo ldconfig10.Install Intel® RealSense™ SDK 2.0
$ cd ~
$ sudo apt update;sudo apt upgrade
$ sudo apt install -y vulkan-utils libvulkan1 libvulkan-dev
# Ubuntu16.04 Only
$ sudo apt install -y mesa-utils* libglu1* libgles2-mesa-dev libopenal-dev gtk+-3.0
# The latest version is unstable
$ cd ~/librealsense/build
$ sudo make uninstall
$ cd ~
$ sudo rm -rf librealsense
$ git clone -b v2.16.5 https://github.com/IntelRealSense/librealsense.git
$ cd ~/librealsense
$ git checkout -b v2.16.5
$ mkdir build;cd build
$ cmake .. -DBUILD_EXAMPLES=true -DCMAKE_BUILD_TYPE=Release
# For RaspberryPi3
$ make -j1
or
# For LaptopPC
$ make -j8
$ sudo make install11.Install Python binding
$ cd ~/librealsense/build
#When using with Python 3.x series
$ cmake .. -DBUILD_PYTHON_BINDINGS=bool:true -DPYTHON_EXECUTABLE=$(which python3)
OR
#When using with Python 2.x series
$ cmake .. -DBUILD_PYTHON_BINDINGS=bool:true -DPYTHON_EXECUTABLE=$(which python)
# For RaspberryPi3
$ make -j1
or
# For LaptopPC
$ make -j8
$ sudo make install12.Update PYTHON_PATH
$ nano ~/.bashrc
export PYTHONPATH=$PYTHONPATH:/usr/local/lib
$ source ~/.bashrc13.RealSense SDK import test
$ python3
Python 3.5.3 (default, Jan 19 2017, 14:11:04)
[GCC 6.3.0 20170124] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import pyrealsense2
>>> exit()14.Installing the OpenGL package for Python
$ sudo apt-get install -y python-opengl
$ sudo -H pip3 install pyopengl
$ sudo -H pip3 install pyopengl_accelerate15.Installation of the imutils package. (For PiCamera)
$ sudo apt-get install -y python3-picamera
$ sudo -H pip3 install imutils --upgrade16.Reduce the SWAP area to the default size (RaspberryPi+Raspbian Stretch / RaspberryPi+Ubuntu Mate Only)
$ sudo nano /etc/dphys-swapfile
CONF_SWAPSIZE=100
$ sudo /etc/init.d/dphys-swapfile restart;swapon -s17.Clone a set of resources
$ git clone https://github.com/PINTO0309/MobileNet-SSD-RealSense.git18.[Optional] Create a RAM disk folder for movie file placement
$ cd /etc
$ sudo cp fstab fstab_org
$ sudo nano fstab
# Mount "/home/pi/movie" on RAM disk.
# Add below.
tmpfs /home/pi/movie tmpfs defaults,size=32m,noatime,mode=0777 0 0
$ sudo reboot1.Execute the following
$ sudo apt update;sudo apt upgrade
$ sudo reboot2.Extend the SWAP area (RaspberryPi+Raspbian Stretch / RaspberryPi+Ubuntu Mate Only)
$ sudo nano /etc/dphys-swapfile
CONF_SWAPSIZE=2048
$ sudo /etc/init.d/dphys-swapfile restart;swapon -s3.Install OpenVINO
$ curl -sc /tmp/cookie "https://drive.google.com/uc?export=download&id=1rBl_3kU4gsx-x2NG2I5uIhvA3fPqm8uE" > /dev/null
$ CODE="$(awk '/_warning_/ {print $NF}' /tmp/cookie)"
$ curl -Lb /tmp/cookie "https://drive.google.com/uc?export=download&confirm=${CODE}&id=1rBl_3kU4gsx-x2NG2I5uIhvA3fPqm8uE" -o l_openvino_toolkit_ie_p_2018.5.445.tgz
$ tar -zxvf l_openvino_toolkit_ie_p_2018.5.445.tgz
$ rm l_openvino_toolkit_ie_p_2018.5.445.tgz
$ sed -i "s|<INSTALLDIR>|$(pwd)/inference_engine_vpu_arm|" inference_engine_vpu_arm/bin/setupvars.sh
$ nano ~/.bashrc
### Add 1 row below
source /home/pi/inference_engine_vpu_arm/bin/setupvars.sh
$ source ~/.bashrc
### Successful if displayed as below
[setupvars.sh] OpenVINO environment initialized
$ sudo usermod -a -G users "$(whoami)"
$ sudo reboot
$ uname -a
Linux raspberrypi 4.14.79-v7+ #1159 SMP Sun Nov 4 17:50:20 GMT 2018 armv7l GNU/Linux
$ sh inference_engine_vpu_arm/install_dependencies/install_NCS_udev_rules.sh
### It is displayed as follows
Update udev rules so that the toolkit can communicate with your neural compute stick
[install_NCS_udev_rules.sh] udev rules installed4.Update udev rule
$ sudo apt install -y git libssl-dev libusb-1.0-0-dev pkg-config libgtk-3-dev
$ sudo apt install -y libglfw3-dev libgl1-mesa-dev libglu1-mesa-dev
$ cd /etc/udev/rules.d/
$ sudo wget https://raw.githubusercontent.com/IntelRealSense/librealsense/master/config/99-realsense-libusb.rules
$ sudo udevadm control --reload-rules && udevadm trigger5.Upgrade to "cmake 3.11.4"
$ cd ~
$ wget https://cmake.org/files/v3.11/cmake-3.11.4.tar.gz
$ tar -zxvf cmake-3.11.4.tar.gz;rm cmake-3.11.4.tar.gz
$ cd cmake-3.11.4
$ ./configure --prefix=/home/pi/cmake-3.11.4
$ make -j1
$ sudo make install
$ export PATH=/home/pi/cmake-3.11.4/bin:$PATH
$ source ~/.bashrc
$ cmake --version
cmake version 3.11.46.Register LD_LIBRARY_PATH
$ nano ~/.bashrc
export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH
$ source ~/.bashrc7.Install Intel® RealSense™ SDK 2.0
$ cd ~
$ sudo apt update;sudo apt upgrade
$ sudo apt install -y vulkan-utils libvulkan1 libvulkan-dev
# Ubuntu16.04 Only
$ sudo apt install -y mesa-utils* libglu1* libgles2-mesa-dev libopenal-dev gtk+-3.0
# The latest version is unstable
$ cd ~/librealsense/build
$ sudo make uninstall
$ cd ~
$ sudo rm -rf librealsense
$ git clone -b v2.16.5 https://github.com/IntelRealSense/librealsense.git
$ cd ~/librealsense
$ git checkout -b v2.16.5
$ mkdir build;cd build
$ cmake .. -DBUILD_EXAMPLES=false -DCMAKE_BUILD_TYPE=Release
# For RaspberryPi3
$ make -j1
or
# For LaptopPC
$ make -j8
$ sudo make install8.Install Python binding
$ cd ~/librealsense/build
#When using with Python 3.x series
$ cmake .. -DBUILD_PYTHON_BINDINGS=bool:true -DPYTHON_EXECUTABLE=$(which python3)
OR
#When using with Python 2.x series
$ cmake .. -DBUILD_PYTHON_BINDINGS=bool:true -DPYTHON_EXECUTABLE=$(which python)
# For RaspberryPi3
$ make -j1
or
# For LaptopPC
$ make -j8
$ sudo make install9.Update PYTHON_PATH
$ nano ~/.bashrc
export PYTHONPATH=$PYTHONPATH:/usr/local/lib
$ source ~/.bashrc10.RealSense SDK import test
$ python3
Python 3.5.3 (default, Jan 19 2017, 14:11:04)
[GCC 6.3.0 20170124] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import pyrealsense2
>>> exit()11.Installing the OpenGL package for Python
$ sudo apt-get install -y python-opengl
$ sudo -H pip3 install pyopengl
$ sudo -H pip3 install pyopengl_accelerate12.Installation of the imutils package. (For PiCamera)
$ sudo apt-get install -y python3-picamera
$ sudo -H pip3 install imutils --upgrade13.Reduce the SWAP area to the default size (RaspberryPi+Raspbian Stretch / RaspberryPi+Ubuntu Mate Only)
$ sudo nano /etc/dphys-swapfile
CONF_SWAPSIZE=100
$ sudo /etc/init.d/dphys-swapfile restart;swapon -s14.Clone a set of resources
$ git clone https://github.com/PINTO0309/MobileNet-SSD-RealSense.git15.[Optional] Create a RAM disk folder for movie file placement
$ cd /etc
$ sudo cp fstab fstab_org
$ sudo nano fstab
# Mount "/home/pi/movie" on RAM disk.
# Add below.
tmpfs /home/pi/movie tmpfs defaults,size=32m,noatime,mode=0777 0 0
$ sudo reboot$ python3 MultiStickSSDwithRealSense.py <option1> <option2> ...
<options>
-grp MVNC graphs Path. (Default=./)
-mod Camera Mode. (0:=RealSense Mode, 1:=USB Camera Mode. Defalut=0)
-wd Width of the frames in the video stream. (USB Camera Mode Only. Default=320)
-ht Height of the frames in the video stream. (USB Camera Mode Only. Default=240)
-tp TransparentMode. (RealSense Mode Only. 0:=No background transparent, 1:=Background transparent. Default=0)
-sd SSDDetectionMode. (0:=Disabled, 1:=Enabled. Default=1)
-fd FaceDetectionMode. (0:=Disabled, 1:=Enabled. Default=0)
-snc stick_num_of_cluster. Number of sticks to be clustered. (0:=Clustering invalid, n:=Number of sticks Default=0)
-csc cluster_switch_cycle. Cycle of switching active cluster. (n:=millisecond Default=10000)
-cst cluster_switch_temperature. Temperature threshold to switch active cluster. (n.n:=temperature(Celsius) Default=65.0)
(Example0) MobileNet-SSD + Neural Compute Stick + RealSense D435 Mode + Syncronous
$ sudo raspi-config
"7.Advanced Options" - "A7 GL Driver" - "G3 Legacy"
$ cd ~/MobileNet-SSD-RealSense
$ python3 SingleStickSSDwithRealSense.py
(Example1) MobileNet-SSD + Neural Compute Stick + RealSense D435 Mode + Asynchronous
$ sudo raspi-config
"7.Advanced Options" - "A7 GL Driver" - "G3 Legacy"
$ cd ~/MobileNet-SSD-RealSense
$ python3 MultiStickSSDwithRealSense.py
(Example2) MobileNet-SSD + Neural Compute Stick + USB Camera Mode + Asynchronous
$ sudo raspi-config
"7.Advanced Options" - "A7 GL Driver" - "G3 Legacy"
$ cd ~/MobileNet-SSD-RealSense
$ python3 MultiStickSSDwithRealSense.py -mod 1 -wd 640 -ht 480
$ python3 MultiStickSSDwithRealSense.py -mod 1 -wd 320 -ht 240
$ sudo raspi-config
"7.Advanced Options" - "A7 GL Driver" - "G3 Legacy"
$ cd ~/MobileNet-SSD-RealSense
$ python3 MultiStickSSDwithRealSense.py -tp 1
(Example4) MobileNet-SSD + FaceDetection + Neural Compute Stick + USB Camera Mode + Asynchronous
$ sudo raspi-config
"7.Advanced Options" - "A7 GL Driver" - "G3 Legacy"
$ cd ~/MobileNet-SSD-RealSense
$ python3 MultiStickSSDwithRealSense.py -mod 1 -wd 640 -ht 480 -fd 1
(Example5) To prevent thermal runaway, simple clustering function (2 Stick = 1 Cluster)
When a certain cycle or constant temperature is reached, the active cluster switches seamlessly automatically.
You must turn on the clustering enable flag.
The default switch period is 10 seconds, the default temperature threshold is 65°C.
The number, cycle, and temperature of sticks constituting one cluster can be specified by the start parameter.
Depending on your environment, please tune to the optimum parameters yourself.
[1] Number of all sticks = 5
[2] stick_num_of_cluster = 2
[3] cluster_switch_cycle = 10sec (10,000millisec)
[4] cluster_switch_temperature = 65.0℃
$ sudo raspi-config
"7.Advanced Options" - "A7 GL Driver" - "G3 Legacy"
$ cd ~/MobileNet-SSD-RealSense
$ python3 MultiStickSSDwithRealSense.py -mod 1 -snc 2 -csc 10000 -cst 65.0
[Simplified drawing of cluster switching]
[Execution log]
(Example6)
$ sudo raspi-config
"7.Advanced Options" - "A7 GL Driver" - "G2 GL (Fake KMS)"
$ realsense-viewer
(Example7)
$ sudo raspi-config
"7.Advanced Options" - "A7 GL Driver" - "G3 Legacy"
$ cd ~/librealsense/wrappers/opencv/build/grabcuts
$ rs-grabcuts
(Example8)
$ sudo raspi-config
"7.Advanced Options" - "A7 GL Driver" - "G3 Legacy"
$ cd ~/librealsense/wrappers/opencv/build/imshow
$ rs-imshow
(Example9) MobileNet-SSD(OpenCV-DNN) + RealSense D435 + Without Neural Compute Stick
$ sudo raspi-config
"7.Advanced Options" - "A7 GL Driver" - "G3 Legacy"
$ cd ~/librealsense/wrappers/opencv/build/dnn
$ rs-dnn
https://github.com/xufeifeiWHU/Mobilenet-v2-on-Movidius-stick.git
$ cd ~/MobileNet-SSD-RealSense
$ mvNCCompile ./caffemodel/MobileNetSSD/deploy.prototxt -w ./caffemodel/MobileNetSSD/MobileNetSSD_deploy.caffemodel -s 12
$ mvNCCompile ./caffemodel/Facedetection/fullface_deploy.prototxt -w ./caffemodel/Facedetection/fullfacedetection.caffemodel -s 12
$ mvNCCompile ./caffemodel/Facedetection/shortface_deploy.prototxt -w ./caffemodel/Facedetection/shortfacedetection.caffemodel -s 12$ cd ~/MobileNet-SSD-RealSense
$ sudo python3 /opt/intel/openvino/deployment_tools/model_optimizer/mo.py \
--input_model caffemodel/MobileNetSSD/MobileNetSSD_deploy.caffemodel \
--input_proto caffemodel/MobileNetSSD/MobileNetSSD_deploy.prototxt \
--data_type FP16 \
--batch 1or
$ cd ~/MobileNet-SSD-RealSense
$ sudo python3 /opt/intel/openvino/deployment_tools/model_optimizer/mo.py \
--input_model caffemodel/MobileNetSSD/MobileNetSSD_deploy.caffemodel \
--input_proto caffemodel/MobileNetSSD/MobileNetSSD_deploy.prototxt \
--data_type FP32 \
--batch 1Construction of learning environment and simple test for model (Ubuntu16.04 x86_64 PC + GPU[NVIDIA Geforce])
1.【Example】 Introduction of NVIDIA-Driver, CUDA and cuDNN to the environment with GPU
$ sudo apt-get remove nvidia-*
$ sudo apt-get remove cuda-*
$ apt search "^nvidia-[0-9]{3}$"
$ sudo apt install cuda-9.0
$ sudo reboot
$ nvidia-smi
### Download cuDNN v7.2.1 NVIDIA Home Page
### libcudnn7_7.2.1.38-1+cuda9.0_amd64.deb
### libcudnn7-dev_7.2.1.38-1+cuda9.0_amd64.deb
### cuda-repo-ubuntu1604-9-0-local_9.0.176-1_amd64.deb
### cuda-repo-ubuntu1604-9-0-local-cublas-performance-update_1.0-1_amd64.deb
### cuda-repo-ubuntu1604-9-0-local-cublas-performance-update-2_1.0-1_amd64.deb
### cuda-repo-ubuntu1604-9-0-local-cublas-performance-update-3_1.0-1_amd64.deb
### cuda-repo-ubuntu1604-9-0-176-local-patch-4_1.0-1_amd64.deb
$ sudo dpkg -i libcudnn7*
$ sudo dpkg -i cuda-repo-ubuntu1604-9-0-local_9.0.176-1_amd64.deb
$ sudo apt-key add /var/cuda-repo-9-0-local/7fa2af80.pub
$ sudo apt update
$ sudo dpkg -i cuda-repo-ubuntu1604-9*
$ sudo apt update
$ rm libcudnn7_7.2.1.38-1+cuda9.0_amd64.deb;rm libcudnn7-dev_7.2.1.38-1+cuda9.0_amd64.deb;rm cuda-repo-ubuntu1604-9-0-local_9.0.176-1_amd64.deb;rm cuda-repo-ubuntu1604-9-0-local-cublas-performance-update_1.0-1_amd64.deb;rm cuda-repo-ubuntu1604-9-0-local-cublas-performance-update-2_1.0-1_amd64.deb;rm cuda-repo-ubuntu1604-9-0-local-cublas-performance-update-3_1.0-1_amd64.deb;rm cuda-repo-ubuntu1604-9-0-176-local-patch-4_1.0-1_amd64.deb
$ echo 'export PATH=/usr/local/cuda-9.0/bin:${PATH}' >> ~/.bashrc
$ echo 'export LD_LIBRARY_PATH=/usr/local/cuda-9.0/lib64:${LD_LIBRARY_PATH}' >> ~/.bashrc
$ source ~/.bashrc
$ sudo ldconfig
$ nvcc -V
$ cd ~;nano cudnn_version.cpp
#include <cudnn.h>
#include <iostream>
int main(int argc, char** argv) {
std::cout << "CUDNN_VERSION: " << CUDNN_VERSION << std::endl;
return 0;
}
$ nvcc cudnn_version.cpp -o cudnn_version
$ ./cudnn_version
$ sudo pip2 uninstall tensorflow-gpu
$ sudo pip2 install tensorflow-gpu==1.10.0
$ sudo pip3 uninstall tensorflow-gpu
$ sudo pip3 install tensorflow-gpu==1.10.0
2.【Example】 Introduction of Caffe to environment with GPU
$ cd ~
$ sudo apt install libopenblas-base libopenblas-dev
$ git clone https://github.com/weiliu89/caffe.git
$ cd caffe
$ git checkout ssd
$ cp Makefile.config.example Makefile.config
$ nano Makefile.config
# cuDNN acceleration switch (uncomment to build with cuDNN).
#USE_CUDNN := 1
↓
# cuDNN acceleration switch (uncomment to build with cuDNN).
USE_CUDNN := 1
# Uncomment if you're using OpenCV 3
# OPENCV_VERSION := 3
↓
# Uncomment if you're using OpenCV 3
OPENCV_VERSION := 3
# CUDA directory contains bin/ and lib/ directories that we need.
CUDA_DIR := /usr/local/cuda
↓
# CUDA directory contains bin/ and lib/ directories that we need.
CUDA_DIR := /usr/local/cuda-9.0
# CUDA architecture setting: going with all of them.
# For CUDA < 6.0, comment the lines after *_35 for compatibility.
CUDA_ARCH := -gencode arch=compute_20,code=sm_20 \
-gencode arch=compute_20,code=sm_21 \
-gencode arch=compute_30,code=sm_30 \
-gencode arch=compute_35,code=sm_35 \
-gencode arch=compute_50,code=sm_50 \
-gencode arch=compute_52,code=sm_52 \
-gencode arch=compute_61,code=sm_61
↓
# CUDA architecture setting: going with all of them.
# For CUDA < 6.0, comment the lines after *_35 for compatibility.
CUDA_ARCH := -gencode arch=compute_30,code=sm_30 \
-gencode arch=compute_35,code=sm_35 \
-gencode arch=compute_50,code=sm_50 \
-gencode arch=compute_52,code=sm_52 \
-gencode arch=compute_61,code=sm_61
# NOTE: this is required only if you will compile the python interface.
# We need to be able to find Python.h and numpy/arrayobject.h.
PYTHON_INCLUDE := /usr/include/python2.7 \
/usr/lib/python2.7/dist-packages/numpy/core/include
↓
# NOTE: this is required only if you will compile the python interface.
# We need to be able to find Python.h and numpy/arrayobject.h.
PYTHON_INCLUDE := /usr/include/python2.7 \
/usr/lib/python2.7/dist-packages/numpy/core/include \
/usr/local/lib/python2.7/dist-packages/numpy/core/include
# Uncomment to support layers written in Python (will link against Python libs)
# WITH_PYTHON_LAYER := 1
↓
# Uncomment to support layers written in Python (will link against Python libs)
WITH_PYTHON_LAYER := 1
# Whatever else you find you need goes here.
INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include
LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib
↓
# Whatever else you find you need goes here.
INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include \
/usr/include/hdf5/serial
LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib \
/usr/lib/x86_64-linux-gnu/hdf5/serial
# Uncomment to use `pkg-config` to specify OpenCV library paths.
# (Usually not necessary -- OpenCV libraries are normally installed in one of the above $LIBRARY_DIRS.)
# USE_PKG_CONFIG := 1
↓
# Uncomment to use `pkg-config` to specify OpenCV library paths.
# (Usually not necessary -- OpenCV libraries are normally installed in one of the above $LIBRARY_DIRS.)
USE_PKG_CONFIG := 1
$ rm -r -f build
$ rm -r -f .build_release
$ make superclean
$ make all -j4
$ make test -j4
$ make distribute -j4
$ export PYTHONPATH=/home/<username>/caffe/python:$PYTHONPATH
$ make py
3.Download of VGG model [My Example CAFFE_ROOT PATH = "/home/<username>/caffe"]
$ export CAFFE_ROOT=/home/<username>/caffe
$ cd $CAFFE_ROOT/models/VGGNet
$ wget http://cs.unc.edu/~wliu/projects/ParseNet/VGG_ILSVRC_16_layers_fc_reduced.caffemodel
4.Download VOC 2007 and VOC 2012 datasets
# Download the data.
$ cd ~;mkdir data;cd data
$ wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar #<--- 1.86GB
$ wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar #<--- 438MB
$ wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar #<--- 430MB
# Extract the data.
$ tar -xvf VOCtrainval_11-May-2012.tar
$ tar -xvf VOCtrainval_06-Nov-2007.tar
$ tar -xvf VOCtest_06-Nov-2007.tar
$ rm VOCtrainval_11-May-2012.tar;rm VOCtrainval_06-Nov-2007.tar;rm VOCtest_06-Nov-2007.tar
5.Generate lmdb file
$ export CAFFE_ROOT=/home/<username>/caffe
$ cd $CAFFE_ROOT
# Create the trainval.txt, test.txt, and test_name_size.txt in $CAFFE_ROOT/data/VOC0712/
$ ./data/VOC0712/create_list.sh
# You can modify the parameters in create_data.sh if needed.
# It will create lmdb files for trainval and test with encoded original image:
# - $HOME/data/VOCdevkit/VOC0712/lmdb/VOC0712_trainval_lmdb
# - $HOME/data/VOCdevkit/VOC0712/lmdb/VOC0712_test_lmdb
# and make soft links at examples/VOC0712/
$ ./data/VOC0712/create_data.sh
6.Execution of learning [My Example environment GPU x1, GeForce GT 650M = RAM:2GB]
Adjust according to the number of GPU
# It will create model definition files and save snapshot models in:
# - $CAFFE_ROOT/models/VGGNet/VOC0712/SSD_300x300/
# and job file, log file, and the python script in:
# - $CAFFE_ROOT/jobs/VGGNet/VOC0712/SSD_300x300/
# and save temporary evaluation results in:
# - $HOME/data/VOCdevkit/results/VOC2007/SSD_300x300/
# It should reach 77.* mAP at 120k iterations.
$ export CAFFE_ROOT=/home/<username>/caffe
$ export PYTHONPATH=/home/<username>/caffe/python:$PYTHONPATH
$ cd $CAFFE_ROOT
$ cp examples/ssd/ssd_pascal.py examples/ssd/BK_ssd_pascal.py
$ nano examples/ssd/ssd_pascal.py
# Solver parameters.
# Defining which GPUs to use.
gpus = "0,1,2,3"
↓
# Solver parameters.
# Defining which GPUs to use.
gpus = "0"
Adjust according to GPU performance (Memory Size) [My Example GeForce GT 650M x1 = RAM:2GB]
# Divide the mini-batch to different GPUs.
batch_size = 32
accum_batch_size = 32
↓
# Divide the mini-batch to different GPUs.
batch_size = 1
accum_batch_size = 1
Execution
- The learned data is generated in "$CAFFE_ROOT/models/VGGNet/VOC0712/SSD_300x300"
- VGG_VOC0712_SSD_300x300_iter_n.caffemodel
- VGG_VOC0712_SSD_300x300_iter_n.solverstate
$ export CAFFE_ROOT=/home/<username>/caffe
$ export PYTHONPATH=/home/<username>/caffe/python:$PYTHONPATH
$ cd $CAFFE_ROOT
$ python examples/ssd/ssd_pascal.py
7.Evaluation of learning data (still image)
$ export CAFFE_ROOT=/home/<username>/caffe
$ export PYTHONPATH=/home/<username>/caffe/python:$PYTHONPATH
$ cd $CAFFE_ROOT
# If you would like to test a model you trained, you can do:
$ python examples/ssd/score_ssd_pascal.py
8.Evaluation of learning data (USB camera)
$ export CAFFE_ROOT=/home/<username>/caffe
$ export PYTHONPATH=/home/<username>/caffe/python:$PYTHONPATH
$ cd $CAFFE_ROOT
# If you would like to attach a webcam to a model you trained, you can do:
$ python examples/ssd/ssd_pascal_webcam.py
https://github.com/movidius/ncappzoo/tree/master/caffe/SSD_MobileNet
https://github.com/FreeApe/VGG-or-MobileNet-SSD
https://github.com/chuanqi305/MobileNet-SSD
https://github.com/avBuffer/MobilenetSSD_caffe
https://github.com/Coldmooon/SSD-on-Custom-Dataset
https://github.com/BVLC/caffe/wiki/Ubuntu-16.04-or-15.10-Installation-Guide#the-gpu-support-prerequisites
https://stackoverflow.com/questions/33962226/common-causes-of-nans-during-training
https://github.com/CongWeilin/mtcnn-caffe
https://github.com/DuinoDu/mtcnn.git
https://www.hackster.io/mjrobot/real-time-face-recognition-an-end-to-end-project-a10826
https://github.com/Mjrovai/OpenCV-Face-Recognition.git
https://github.com/sgxu/face-detection-based-on-caffe.git
https://github.com/RiweiChen/DeepFace.git
https://github.com/KatsunoriWa/eval_faceDetectors
https://github.com/BeloborodovDS/MobilenetSSDFace
https://www.pyimagesearch.com/2018/09/03/semantic-segmentation-with-opencv-and-deep-learning/
https://github.com/TimoSaemann/ENet/tree/master/Tutorial
https://blog.amedama.jp/entry/2017/04/03/235901
https://github.com/NVIDIA/nvidia-docker
https://hub.docker.com/r/nvidia/cuda/
https://www.dlology.com/blog/how-to-run-keras-model-on-movidius-neural-compute-stick/
https://ncsforum.movidius.com/discussion/1106/ncs-temperature-issue
https://github.com/opencv/opencv/wiki/Intel%27s-Deep-Learning-Inference-Engine-backend
https://github.com/opencv/opencv/wiki/Intel%27s-Deep-Learning-Inference-Engine-backend#raspbian-stretch
https://github.com/skhameneh/OpenVINO-ARM64